Site Link Analyzer

Analyze site links

Find and fix broken links across the site. The work splits in two:

• scripts/check-links.mjs (run via bun run lint:links) is the deterministic source of truth. It extracts links and validates them, prints a report, and exits non-zero when any link is broken. It never edits files.
• You are the judgment layer: read the checker's JSON, apply the confident fixes, leave the ambiguous ones for the human, and re-run the checker to confirm green before you call the job done.

The checker re-derives the valid route set from the filesystem using Next.js App Router conventions (app/**/page.{js,jsx,ts,tsx,md,mdx}, with route groups, parallel slots, private folders, and [param]/[...catch-all] segments) and treats files under public/ as valid assets, so its results match what the site actually serves. It has no dependencies — it runs on plain Node or Bun.

Before you start: branch

Per .agents/rules/branching-rule.mdc, never commit to main. If the current branch is main, create a branch first (offer a suggested name such as fix/broken-links and a "provide your own" option) before editing any files.

Workflow

- [ ] 1. Run `bun run lint:links -- --json` and read the findings.
- [ ] 2. Triage each finding by status + category.
- [ ] 3. Apply the CONFIDENT auto-fixes (rules below). Leave report-only items.
- [ ] 4. Re-run `bun run lint:links` and confirm it exits 0 (no broken links).
- [ ] 5. Summarize: what you fixed, what needs a human decision, coverage gaps.
- [ ] 6. (Optional) External sweep: `bun run lint:links -- --external` — report only.

How to run

Exit codes: 0 = no broken links, 1 = broken links found, 2 = usage/internal error.

External checking is off by default and never gates CI (rate limits, transient outages, and CI-IP blocking make it non-deterministic). The deterministic internal/anchor/relative checks are the gate; the scheduled .github/workflows/link-check-external.yml runs the --external sweep weekly, non-blocking.

Reading the JSON

{
  "summary": { "filesScanned", "linksChecked", "ok", "broken", "warning", "skipped", "byCategory", "external" },
  "findings": [
    {
      "file": "app/docs/installing/page.js",
      "line": 65, "column": 27,
      "rawLink": "/docs/codin",
      "category": "internal",        // internal | anchor | relative | external | dynamic-skipped | malformed
      "status": "broken",            // ok | broken | warning | skipped
      "reason": "no route resolves \"/docs/codin\"",
      "suggestion": "/docs/coding",
      "suggestionConfidence": "high" // high | medium | low | none — drives the fix decision
    }
  ]
}

suggestionConfidence is the signal for whether to auto-fix. Only broken findings fail the gate; warning and skipped never do.

Auto-fix decision rules

Confident — apply the suggestion, then re-run to confirm:

• category: "malformed" (link target wrapped in backticks) — strip the backticks.
• category: "internal", status: "broken", suggestionConfidence: "high" with a single unambiguous match (including trailing-slash / case normalization, or an obvious typo like /docs/codin → /docs/coding).

Report-only — never auto-edit; surface for a human:

• category: "external", status: "broken" — could be the remote's fault or a transient outage; confirm before touching content.
• category: "relative" — points at a repo file (e.g. deploy/README.md); intended for GitHub, not the site. Leave it unless the human says otherwise.
• suggestionConfidence: "medium" | "low" | "none", or multiple tied candidates — present the options and ask.
• category: "dynamic-skipped" and category: "anchor" — nothing to fix; the checker can't verify these (see Coverage).

Re-run to verify

After applying fixes, run bun run lint:links again and show the result. Do not declare success until it exits 0, or until every remaining finding is a deliberate warning/skipped you've explained. Never claim green without the re-run.

Coverage & limits

• Dynamic links are skipped by design. href={expr} / src={expr} expressions (including the shared link constants used as href={GITHUB_URL}) can't be statically resolved and are reported as dynamic-skipped. The URLs behind those constants are still checked separately: the checker reads external URL string literals (e.g. export const GITHUB_URL = "https://…" in components/links.js), so the --external sweep covers them. Coverage is intentionally partial — check the dynamic-skipped count so nobody assumes 100%.
• Anchors are not statically verifiable. App Router JSX/MDX headings get no auto-generated ids (there is no rehype-slug step), so a #fragment is reported skipped. The path portion of /path#fragment is still validated.
• Dynamic routes (app/blog/[slug]/page.js, [...catch-all]) are recognized as valid route templates; concrete /blog/<slug> links match them. A path that matches no static route and no template still fails.
• Static assets under public/ (e.g. /images/logo.png, /favicon.ico) are treated as valid.
• Only href/src string-literal attributes, external URL string literals, and markdown links are scanned. Other string literals (class names, metadata keys, etc.) are intentionally ignored to avoid false positives.

Maintaining the checker

When routing or asset conventions change (a new App Router segment style, a new content directory, a new link-constant convention), update scripts/check-links.mjs: the valid-target model is built in buildModel() / routeFromAppFile(), link extraction lives in scanCodeFile / scanMarkdownFile, and classification in validateTarget. Keep route derivation in sync with Next.js App Router rules.

When you add or change this skill, update the Skills table in AGENTS.md.